CN106875425A - A kind of multi-target tracking system and implementation method based on deep learning - Google Patents
A kind of multi-target tracking system and implementation method based on deep learning Download PDFInfo
- Publication number
- CN106875425A CN106875425A CN201710053918.6A CN201710053918A CN106875425A CN 106875425 A CN106875425 A CN 106875425A CN 201710053918 A CN201710053918 A CN 201710053918A CN 106875425 A CN106875425 A CN 106875425A
- Authority
- CN
- China
- Prior art keywords
- target
- frame
- tracking
- trail
- followed
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of multi-target tracking system and implementation method based on deep learning, method includes：The target location of the first frame is obtained by target detection, multiple target to be followed the trail of is added in tracking queue, input next frame picture simultaneously travels through the tracking queue, target position in the next frame is obtained, above-mentioned target is being obtained after the position of next frame, whether the target frames out by threshold decision, if not, a target detection then is called every an anchor-frame, the result of target detection and the result followed the trail of are calculated into IOU hands over and compare, if IOU<0.1, then it is assumed that new target adds screen, by target addition tracking queue；If IOU>0.5, then the frame followed the trail of is substituted using the frame of target detection, carry out position correction；Continuation is tracked to target.The present invention improves training method by well-designed network structure, while height follows the trail of precision, significantly improves the speed of tracking, reduces the redundancy of network, reduces the size of model.
Description
Technical field
The present invention relates to image processing field, more particularly to a kind of multi-target tracking system and realization based on deep learning
Method, realizes the problem of and fast track accurate for multiple target.
Background technology
Motion target tracking is exactly in a continuous videos sequence, fortune interested to be found in each frame monitored picture
Moving-target (for example, vehicle, pedestrian, animal etc.).Tracking can be roughly divided into following steps：
1) effective description of target；The tracking process of target as target detection, it is necessary to carry out effective description to it,
That is, it needs to extract clarification of objective such that it is able to express the target；In general, we can be by the edge of image, wheel
Exterior feature, shape, texture, region, histogram, moment characteristics, conversion coefficient etc. carry out clarification of objective description；
2) similarity measurement is calculated；Conventional method has：Euclidean distance, mahalanobis distance, chessboard distance, Weighted distance, phase
Like coefficient, coefficient correlation etc.；
3) target area search matching；If all targets to occurring in scene all carry out feature extraction, similitude meter
Calculate, then the amount of calculation spent by system operation is very big.So, at present may to moving target generally by the way of certain
The region of appearance is estimated, so as to reduce redundancy, accelerates the speed of target following；Common prediction algorithm has：Kalman is filtered
Ripple, particle filter, average drifting etc..
Based on above-mentioned, the target tracking algorithm of motion target tracking generally comprises the tracking based on active profile, based on spy
Tracking, the tracking based on region and the tracking based on model levied.The precision and robustness of track algorithm largely take
Certainly in expression and the definition of similarity measurement to moving target, the real-time of track algorithm depends on matching search strategy and filter
Ripple prediction algorithm.
It is the deformable curve defined in image area i.e. Snake curves, by its energy based on the tracking of active profile
The minimum of flow function, it is consistent with objective contour that dynamic outline progressively adjusts own form.Snake technologies can be processed arbitrarily
Any deformation of shaped objects, will split the object boundary for obtaining as the original template for tracking it is then determined that characterizing object first
The object function of real border, and by reducing target function value, initial profile is gradually moved to the real border of object.Base
It is not only to consider the half-tone information from image in the advantage of active Contour extraction, and considers the geological information of overall profile,
Enhance the reliability of tracking.Because tracking process is actually the searching process of solution, the amount of calculation brought is than larger, Er Qieyou
In the blindness of Snake models, during for the quick object for moving or larger deformation, tracking effect is not ideal enough.
The tracking of feature based, its global feature for not considering moving target, only by some notable spies of target image
Levy to be tracked.It is assumed that moving target can be expressed by only characteristic set, search the corresponding characteristic set and just recognize
For moving target has been gone up in tracking.Mainly include two aspects of feature extraction and characteristic matching：The purpose of feature extraction is to carry out frame
Between target signature matching, and target is tracked with Optimum Matching.The track algorithm of common feature based matching has based on two
The tracking of value target image matching, the tracking matched based on Edge Feature Matching or Corner Feature, based on target gray feature
The tracking of matching, the tracking based on color of object characteristic matching etc..The track algorithm of feature based is for image blurring, noise etc.
Compare sensitive, the extraction effect of characteristics of image also relies on the setting of various extraction operators and its parameter, additionally, continuous interframe
The also more difficult determination of feature corresponding relation, especially when the number of features of each two field picture is inconsistent, there is missing inspection, feature increase or
Situations such as reduction.
Tracking based on region, by obtaining the template comprising target, the template can be obtained or advance by image segmentation
Artificial to determine, template is usually the rectangle slightly larger than target, or irregular shape；In sequence image, calculated with correlation
Method tracks target.Its shortcoming is time-consuming first, and when region of search is larger, situation is especially serious；Secondly, algorithm requirement target becomes
Shape less, and can not have and block greatly very much, and otherwise related precise decreasing can cause the loss of target.In recent years, to based on region
It is situation when how processing template changes that tracking is paid close attention to more, and this change is caused by moving object attitude change
, if the attitudes vibration of the correctly predicted target of energy, is capable of achieving tenacious tracking.
Tracking based on model, is to set up model to tracked target by certain priori, then by matching
Tracking target carries out the real-time update of model.For rigid-object, the conversion of its motion state mainly translation, rotation etc.,
Target following can be realized using the method.But what is tracked in practical application is not only rigid body, also most right and wrong
Rigid body, the definite geometrical model of target is not readily available.This method is difficult to be influenceed by observation visual angle, with stronger robust
Property, Model Matching tracking accuracy is high, is suitable for the various motion changes of maneuvering target, strong antijamming capability, but due to calculating point
Analysis is complicated, arithmetic speed is slow, and the renewal of model is complex, and real-time is poor.It is Model Matching energy accurately to set up motion model
No successful key.
It is many to there is redundancy in present target tracking algorithm network, and speed is slow, and model is big, it is difficult to practical, it is impossible to chase after in real time
The problems such as track.And the concept of deep learning comes from the research of artificial neural network, deep learning is formed by combining low-level feature
More abstract high-rise expression attribute classification or feature, is represented with the distributed nature for finding data.
The content of the invention
The technical problem to be solved in the present invention is to provide by object detection one each target object of reference frame of acquisition
Accurate location, and that can carry out prolonged tracking to each target object on the basis of target initial position is him.
Above-mentioned technical problem is solved, the invention provides a kind of multi-target tracking method based on deep learning, including such as
Lower step：
The target location of the first frame is obtained by target detection, multiple target to be followed the trail of is added in tracking queue,
Input next frame picture simultaneously travels through the tracking queue, obtains target position in the next frame,
Specifically, when the target location of next frame is predicted, system and the picture and picture of previous frame are saved
The position of middle target, volume and neutral net can predict target in next frame by contrasting the difference of previous frame and next frame
Position；
Above-mentioned target is being obtained after the position of next frame, whether the target frames out by threshold decision,
If it is not, then calling a target detection every an anchor-frame, the result of target detection is calculated with the result followed the trail of
IOU is handed over and compared,
If the IOU of a result of target detection and all targets followed the trail of<0.1, then it is assumed that new target adds screen
Curtain, by target addition tracking queue；
If IOU>0.5, then the frame followed the trail of is substituted using the frame of target detection, carry out position correction；
Continuation is tracked to target.
Further, method also includes following pre-training process：
By this two pictures by change of scale to same yardstick, two kinds of picture conducts of similar adjacent video frames are obtained
Training picture, pre-training is carried out to network.
Further, using the picture of ILSVRC contest target detections DET as above-mentioned training picture, based on ImageNet
The image data base of target detection DET this task, ImageNet has 5 tasks, the different view data of each task correspondence
Storehouse.
Further, method also includes following training process：
Pass through the twin network extraction picture feature of parameter identical after two pictures are pre-processed first；
Secondly, the twin network by it is dense->Sparse->Dense convolutional neural networks extract picture feature；
Then, two features are subtracted each other the feature as fusion, this feature is returned by full articulamentum again then
Obtain the position of target frame.
Further, using CRELU joint amendment linear units in the characteristic extraction procedure of convolutional neural networks.
Further, the target position of first frame is detected using the target detection technique based on faster-rcnn frameworks
Put.
Further, the anchor-frame is 10.
Further, judge that whether the target threshold condition that frames out is：
h/w>threshold1、w/h>threshold2、|x1|/W<threshold3、|W-x2|/W<threshold3、|
y1|/H<threshold4、|H-y1|/H<Any in the threshold condition of threshold4,
Wherein, threshold represents threshold value, and h and w is respectively the height and width of object, and H and W is respectively the height and width of frame,
(x1, y1) is the point coordinates in the target upper left corner, and (x2, y2) is the point coordinates in the target lower right corner.
Further, if multi-target tracking is face tracking, threshold1=threshold2=2 is set,
Threshold3=threshold4=0.02.
A kind of multi-target tracking system based on deep learning is additionally provided based on the invention described above, including：
Training unit, is used to the target location that target detection obtains the first frame, multiple target to be followed the trail of is added to
In tracking queue；
Detection unit, is used to the next frame picture to being input into and travels through the tracking queue, obtains the target in next frame
In position；And when target does not frame out, a target detection is called every an anchor-frame；
Tracing unit, to obtaining above-mentioned target after the position of next frame, by the threshold decision target whether
The finger target object that frame out of frameing out has not been suffered in picture and has frameed out finger target object not in picture
；
Threshold cell, is used to the result of target detection and the result followed the trail of calculating IOU are handed over and compared, if target detection
The IOU of one result and all targets followed the trail of<0.1, then it is assumed that new target adds screen, the target is added and follows the trail of queue
In；If IOU>0.5, then the frame followed the trail of is substituted using the frame of target detection, carry out position correction
Beneficial effects of the present invention：
In the present invention, by well-designed network structure, training method is improved, while height follows the trail of precision, significantly
The speed of tracing algorithm is improve, the redundancy of network is reduced, the size of model is reduced, and object tracking algorithm is subject to
It is practical.
Additionally, the present invention is also equipped with following advantage：1) speed is fast, and the tracking to single goal 120fps is reached on i5cpu
Speed, but the 25fps of the video frame rate of current main-stream, therefore the method for the present invention can carry out reality to the target object in video
When follow the trail of.2) model is small, and by simulated experiment, model size is 5.5M, compared to the deep learning tracing algorithm of existing main flow
Model it is small more than 50 times.3) accuracy rate is higher, and the present invention is tested in object data set ALOV and human face data collection 300VW
Card.4) multi-target tracking is realized, the tracing model of current main flow is difficult to be tracked multiple target, the tracing system in the present invention
Multiple targets can be in real time tracked.
Brief description of the drawings
Fig. 1 is the method flow schematic diagram in the present invention；
Fig. 2 is the system structure diagram in the present invention；
Fig. 3 is training process schematic diagram；
Fig. 4 is to implement schematic flow sheet in one embodiment of the present invention；
Fig. 5 (a)-Fig. 5 (c) is that result schematic diagram is followed the trail of in the simulation in the present invention.
Specific embodiment
The principle of the disclosure is described referring now to some example embodiments.It is appreciated that these embodiments are merely for saying
It is bright and help it will be understood by those skilled in the art that with the purpose of the embodiment disclosure and describe, rather than advise model of this disclosure
Any limitation enclosed.Content of this disclosure described here can be implemented in the various modes outside mode described below.
As described herein, term " including " and its various variants be construed as open-ended term, it means that " bag
Include but be not limited to ".Term "based" is construed as " being based at least partially on ".Term " one embodiment " it is understood that
It is " at least one embodiment ".Term " another embodiment " is construed as " at least one other embodiment ".
It is appreciated that the concept being defined as follows in this application：
The CRELU refers to joint amendment linear unit.
The parameter sharing refers to a kind of algorithm of characteristic similarity study.
The Fusion Features are included but is not limited to, and eigenmatrix is merged, so as to multiple features are become into one more
Effective fusion feature.
The IOU refers to hand over and compare, i.e., two intersection of sets collection are divided by two union of sets collection
Training includes but is not limited to be trained offline using mass data under the line, and model is carried out after training
The renewal of model parameter is not carried out when test.The concept trained in the concept and line trained under line is relative.
The twin network refers to, two completely identical in structure network structures.
It is as shown in Figure 1 the method flow schematic diagram in the present invention, a kind of multi-target tracking method based on deep learning,
Comprise the following steps：
Step S100 obtains the target location of the first frame by target detection, and multiple target to be followed the trail of is added into tracking team
In row,
Step S101 is input into next frame picture and travels through the tracking queue, obtains target position in the next frame,
Step S102 is obtaining above-mentioned target after the position of next frame, and by threshold decision, whether the target leaves screen
Curtain,
Step S103 if it is not, then call a target detection every an anchor-frame, by the result of target detection with follow the trail of
Result calculates IOU and hands over and compare,
If the IOU of a result of step S104 target detections and all targets followed the trail of<0.1, then it is assumed that new target
Screen is added, by target addition tracking queue；
If step S105 IOU>0.5, then the frame followed the trail of is substituted using the frame of target detection, carry out position correction；
Continue to be tracked target after terminating above-mentioned steps.
Used as preferred in the present embodiment, the method for the present embodiment also includes following pre-training process：By this two pictures
By change of scale to same yardstick, two kinds of pictures of similar adjacent video frames are obtained as training picture, network is carried out
Pre-training.And the picture of ILSVRC contest target detections DET is used as above-mentioned training picture.
Used as preferred in the present embodiment, method also includes following training process：
Pass through the twin network extraction picture feature of parameter identical after two pictures are pre-processed first；First by two
By the twin network extraction picture feature of parameter identical after picture pretreatment, the output of the length representative convolution of figure center is led to
Road number.
Secondly, the twin network by it is dense->Sparse->Dense convolutional neural networks extract picture feature；Convolution
The network for carrying characteristic is a structure of sparse-dense-sparse, i.e., first extract feature with the few convolution of port number,
Connecing the convolution more than a port number again carries out denseization to feature, then to connect a few convolution of port number dilute to the feature of denseization
Thinization, by above-mentioned feature, can reduce redundancy.
Then, two features are subtracted each other the feature as fusion, this feature is returned by full articulamentum again then
Obtain the position of target frame.
As preferred in the present embodiment, amendment is combined using CRELU in the characteristic extraction procedure of convolutional neural networks
Linear unit.
As preferred in the present embodiment, using the target detection technique detection based on faster-rcnn frameworks described the
The target location of one frame.
Used as preferred in the present embodiment, the anchor-frame is 10.
As preferred in the present embodiment, judge that whether the target threshold condition that frames out is：
h/w>threshold1、w/h>threshold2、|x1|/W<threshold3、|W-x2|/W<threshold3、|
y1|/H<threshold4、|H-y1|/H<Any in the threshold condition of threshold4,
Wherein, threshold represents threshold value, and h and w is respectively the height and width of object, and H and W is respectively the height and width of frame,
(x1, y1) is the point coordinates in the target upper left corner, and (x2, y2) is the point coordinates in the target lower right corner.
As preferred in the present embodiment, if multi-target tracking is face tracking, threshold1=is set
Threshold2=2, threshold3=threshold4=0.02.
Method in the present embodiment, by well-designed network structure, improves training method, and the same of precision is followed the trail of in height
When, the speed of tracing algorithm is significantly improved, the redundancy of network is reduced, reduce the size of model, and by object tracking
Algorithm is subject to practicality.
Fig. 2 is the system structure diagram in the present invention, a kind of multiple target based on deep learning in the present embodiment
Tracing system 100, including：
Training unit 1, is used to the target location that target detection obtains the first frame, multiple target to be followed the trail of is added to
In tracking queue；
Detection unit 2, is used to the next frame picture to being input into and travels through the tracking queue, obtains the target in next frame
In position；And when target does not frame out, a target detection is called every an anchor-frame；
Tracing unit 3, to obtaining above-mentioned target after the position of next frame, by the threshold decision target whether
The finger target object that frame out of frameing out has not been suffered in picture and has frameed out finger target object not in picture
；
Threshold cell 4, is used to the result of target detection and the result followed the trail of calculating IOU are handed over and compared, if target detection
A result and follow the trail of all targets IOU<0.1, then it is assumed that new target adds screen, the target is added and follows the trail of team
In row；If IOU>0.5, then the frame followed the trail of is substituted using the frame of target detection, carry out position correction.
Understood according to above-mentioned, in the present embodiment on the basis of adjacent two frame is obtained, carried by convolutional neural networks (CNN)
Picture feature is taken, the feature of adjacent two frame is merged.Simultaneously by feature it is dense->Sparse->Dense network structure sets
Meter mode, and CRELU corrects linear unit in the characteristic extraction procedure of convolutional neural networks using CRELU joints.Protecting
While holding high-accuracy, model size and capacity are reduced so that model can be used in embedded device.
Specifically, in practical scene, an accurate location for each target object of reference frame is obtained by object detection,
Tracing system in the present embodiment can be for a long time followed the trail of each target object on the basis of target initial position.
Principles illustrated of the invention：
First, pre-training
In training process, video data collects relatively difficult, and the network convergence of a random initializtion is slow, so looking for
It is critically important for the algorithm of object tracking to a good pre-training method.
It may be speculated that position coordinates of the object between adjacent two frame meets certain regularity of distribution.
If (c'x,c'y) coordinate of target's center's point, (c in present framex,cy) it is the coordinate of target's center's point in previous frame,
W and h are respectively the wide and height of the rectangle frame of target previous frame.W' and h' is the wide and height of the rectangle frame of present frame.△ x and △ y
Centered on put changed factor in x directions and y directions.
c'x=cx+wΔx (1)
c'y=cy+hΔy (2)
W '=w γw (3)
H'=h γh (4)
Research shows that △ x, △ y, w' and h' meet laplacian distribution：
Using this rule, a region of target position periphery to a static images, can be intercepted, be become
Shape treatment, including zoom, affine transformation etc. obtain the picture after a deformation, and this two pictures is passed through into change of scale
To same yardstick.Two kinds of pictures of similar adjacent video frames are obtained, as training picture.
With the picture of ILSVRC contest target detections DET as training picture, pre-training is carried out to network.
2nd, training process
Be as shown in Figure 3 training process schematic diagram training process flow chart specific as follows shown in：
By the twin network extraction picture feature of parameter identical after two pictures are pre-processed first, the length of figure center
Degree represents the output channel number of convolution.It can be seen that, the network that convolution carries characteristic is a sparse-dense-sparse
Structure, i.e., first extract feature with the few convolution of port number, then the convolution connect more than a port number carries out denseization to feature, then
Feature rarefaction of the few convolution of port number to denseization is connect, most important feature is proposed, redundancy is reduced.Conventional part will
Two features are subtracted each other the feature as fusion after feature extraction out, this feature is being carried out by full articulamentum to return
To the position of the frame of target.
3rd, system flow：As shown in figure 4,
Step S1 target detections obtain the target location of the first frame
Step S2 is put into queue by target is followed the trail of
Step S3 multi-target trackings
Whether step S4 targets frame out, if then entering step S5, if then entering step S6
Step S5 is removed and is followed the trail of queue
Every 10 frames of step S6 call a target detection
Step S7 IOU<0.1, if then entering step S10
Step S8 IOU>0.5, if then entering step S12
Step S9 corrects the result followed the trail of with object detection results
Step S10 detects fresh target
Step S11 adds object queue
Step S12 occurs following the trail of or detection is abnormal
Step S13 continues to follow the trail of
First, the target location of the first frame is detected using the target detection technique based on faster-rcnn frameworks, by multiple
Target to be followed the trail of is added in tracking queue.Input next frame picture, then traversal tracking queue, for each tracking target
Tracing algorithm is called to obtain target position in the next frame.The target is obtained after the position of next frame, by threshold value
Judge whether the target frames out.If the target have left screen, the target is removed and follows the trail of object queue.
Every 10 frames, a target detection is called, the result of target detection is calculated into IOU with the result followed the trail of, if mesh
Mark certain result of detection and the IOU of all targets followed the trail of<0.1, then it is assumed that new target adds screen, the target is added
In entering to follow the trail of queue.If IOU>0.5, the frame followed the trail of is substituted using the frame of target detection, carry out position correction.
Judge that the condition whether target frames out is (meeting either condition)：
h/w>threshold1 (6)
w/h>threshold2 (7)
|x1|/W<threshold3 (8)
|W-x2|/W<threshold3 (9)
|y1|/H<threshold4 (10)
|H-y1|/H<threshold4 (11)
H and w are respectively the height and width of object, and H and W is respectively the height and width of frame, and (x1, y1) sits for the point in the target upper left corner
Mark, (x2, y2) is the point coordinates in the target lower right corner, is collected in the tracing system of different objects, and threshold can take different
Value, in face tracking system, is set to threshold1=threshold2=2, threshold3=threshold4=
0.02。
Fig. 5 (a)-Fig. 5 (c) is that result schematic diagram is followed the trail of in the simulation in the present invention, is input into one section of pedestrian's video, is first according to
Video is cut frame sequence by 25fps.Video is named according to frame number.First the picture to the first frame detects, such as Fig. 5
A ()-Fig. 5 (c) is shown one by one, it can be seen that the first frame detects four faces.Using this four faces as target is followed the trail of, will
Target is followed the trail of for this four to be added in tracking queue.Input next frame, tracing algorithm carries out real-time tracing against four targets.Often
A Face datection is called every 10 frames, fresh target has been detected whether, if fresh target, fresh target tracking queue is added to
In；If without fresh target, the frame according to Face datection is corrected to the frame followed the trail of.
It should be appreciated that each several part of the invention can be realized with hardware, software, firmware or combinations thereof.Above-mentioned
In implementation method, the software that multiple steps or method can in memory and by suitable instruction execution system be performed with storage
Or firmware is realized.If for example, realized with hardware, and in another embodiment, can be with well known in the art
Any one of row technology or their combination are realized：With the logic gates for realizing logic function to data-signal
Discrete logic, the application specific integrated circuit with suitable combinational logic gate circuit, programmable gate array (PGA), scene
Programmable gate array (FPGA) etc..
In the description of this specification, reference term " one embodiment ", " some embodiments ", " example ", " specifically show
The description of example " or " some examples " etc. means to combine specific features, structure, material or spy that the embodiment or example are described
Point is contained at least one embodiment of the invention or example.In this manual, to the schematic representation of above-mentioned term not
Necessarily refer to identical embodiment or example.And, the specific features of description, structure, material or feature can be any
One or more embodiments or example in combine in an appropriate manner.
In general, the various embodiments of the disclosure can be with hardware or special circuit, software, logic or its any combination
Implement.Some aspects can be implemented with hardware, and some other aspect can be with firmware or software implementation, and the firmware or software can
With by controller, microprocessor or other computing devices.Although the various aspects of the disclosure be shown and described as block diagram,
Flow chart is represented using some other drawing, but it is understood that frame described herein, equipment, system, techniques or methods can
With in a non limiting manner with hardware, software, firmware, special circuit or logic, common hardware or controller or other calculating
Equipment or some of combination are implemented.
In addition, although operation is described with particular order, but this is understood not to require this generic operation with shown suitable
Sequence is performed or performed with generic sequence, or requires that all shown operations are performed to realize expected result.In some feelings
Under shape, multitask or parallel processing can be favourable.Similarly, although the details of some specific implementations is superincumbent to beg for
In by comprising, but these are not necessarily to be construed as any limitation of scope of this disclosure, but the description of feature is only pin
To specific embodiment.Some features described in some separate embodiments can also in combination be held in single embodiment
OK.Mutually oppose, the various features described in single embodiment can also be implemented separately or to appoint in various embodiments
The mode of what suitable sub-portfolio is implemented.
Claims (10)
1. a kind of multi-target tracking method based on deep learning, it is characterised in that comprise the following steps：
The target location of the first frame is obtained by target detection, multiple target to be followed the trail of is added in tracking queue,
Input next frame picture simultaneously travels through the tracking queue, obtains target position in the next frame,
Above-mentioned target is being obtained after the position of next frame, whether the target frames out by threshold decision,
If it is not, then calling a target detection every an anchor-frame, the result of target detection and the result followed the trail of are calculated into IOU hands over
And compare,
If the IOU of a result of target detection and all targets followed the trail of<0.1, then it is assumed that new target adds screen, will
The target is added in tracking queue；
If IOU>0.5, then the frame followed the trail of is substituted using the frame of target detection, carry out position correction；
Continuation is tracked to target.
2. multi-target tracking method according to claim 1, it is characterised in that also including following pre-training process：
By this two pictures by change of scale to same yardstick, two kinds of pictures of similar adjacent video frames are obtained as training
Picture, pre-training is carried out to network.
3. multi-target tracking method according to claim 2, it is characterised in that using ILSVRC contest target detections DET
Picture as above-mentioned training picture.
4. multi-target tracking method according to claim 2, it is characterised in that also including following training process：
Pass through the twin network extraction picture feature of parameter identical after two pictures are pre-processed first；
Secondly, the twin network by it is dense->Sparse->Dense convolutional neural networks extract picture feature；
Then, two features are subtracted each other the feature as fusion, this feature is returned by full articulamentum again then
The position of target frame.
5. multi-target tracking method according to claim 1, it is characterised in that in the feature extraction of convolutional neural networks
Using CRELU joint amendment linear units in journey.
6. multi-target tracking method according to claim 1, it is characterised in that using based on faster-rcnn frameworks
Target detection technique detects the target location of first frame.
7. multi-target tracking method according to claim 1, it is characterised in that the anchor-frame is 10.
8. multi-target tracking method according to claim 1, it is characterised in that judge whether target frames out threshold value bar
Part is：
h/w>threshold1、w/h>threshold2、|x1|/W<threshold3、|W-x2|/W<threshold3、|y1|/
H<threshold4、|H-y1|/H<Any in the threshold condition of threshold4,
Wherein, threshold represents threshold value, and h and w is respectively the height and width of object, and H and W is respectively the height and width of frame, (x1,
Y1) it is the point coordinates in the target upper left corner, (x2, y2) is the point coordinates in the target lower right corner.
9. multi-target tracking method according to claim 7, it is characterised in that if multi-target tracking is face tracking,
If threshold1=threshold2=2, threshold3=threshold4=0.02.
10. a kind of multi-target tracking system based on deep learning, it is characterised in that including：
Training unit, is used to carry out the target location that target detection obtains the first frame, multiple target to be followed the trail of being added into tracking
In queue；
Detection unit, is used to the next frame picture to being input into and travels through the tracking queue, obtains the target in the next frame
Position；And when target does not frame out, a target detection is called every an anchor-frame；
Tracing unit, to obtain above-mentioned target after the position of next frame, by threshold decision, whether the target is left
Screen is to frame out to refer to that target object has not suffered the finger target object that frames out and do not suffered in picture in picture；
Threshold cell, is used to the result of target detection and the result followed the trail of calculating IOU are handed over and compared, if a knot of target detection
Fruit and the IOU of all targets followed the trail of<0.1, then it is assumed that new target adds screen, by target addition tracking queue；
If IOU>0.5, then the frame followed the trail of is substituted using the frame of target detection, carry out position correction.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710053918.6A CN106875425A (en) | 2017-01-22 | 2017-01-22 | A kind of multi-target tracking system and implementation method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710053918.6A CN106875425A (en) | 2017-01-22 | 2017-01-22 | A kind of multi-target tracking system and implementation method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106875425A true CN106875425A (en) | 2017-06-20 |
Family
ID=59158713
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710053918.6A Pending CN106875425A (en) | 2017-01-22 | 2017-01-22 | A kind of multi-target tracking system and implementation method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106875425A (en) |
Cited By (47)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107274433A (en) * | 2017-06-21 | 2017-10-20 | 吉林大学 | Method for tracking target, device and storage medium based on deep learning |
| CN107330920A (en) * | 2017-06-28 | 2017-11-07 | 华中科技大学 | A kind of monitor video multi-target tracking method based on deep learning |
| CN107730538A (en) * | 2017-10-11 | 2018-02-23 | 恩泊泰（天津）科技有限公司 | A kind of method and device of the multiple target tracking based on image |
| CN107730536A (en) * | 2017-09-15 | 2018-02-23 | 北京飞搜科技有限公司 | A kind of high speed correlation filtering object tracking method based on depth characteristic |
| CN107845105A (en) * | 2017-10-24 | 2018-03-27 | 深圳市圆周率软件科技有限责任公司 | A kind of monitoring method, smart machine and storage medium based on the linkage of panorama rifle ball |
| CN107886120A (en) * | 2017-11-03 | 2018-04-06 | 北京清瑞维航技术发展有限公司 | Method and apparatus for target detection tracking |
| CN107993250A (en) * | 2017-09-12 | 2018-05-04 | 北京飞搜科技有限公司 | A kind of fast multi-target pedestrian tracking and analysis method and its intelligent apparatus |
| CN107992826A (en) * | 2017-12-01 | 2018-05-04 | 广州优亿信息科技有限公司 | A kind of people stream detecting method based on the twin network of depth |
| CN108090918A (en) * | 2018-02-12 | 2018-05-29 | 天津天地伟业信息系统集成有限公司 | A kind of Real-time Human Face Tracking based on the twin network of the full convolution of depth |
| CN108196680A (en) * | 2018-01-25 | 2018-06-22 | 盛视科技股份有限公司 | Robot vision following method based on human body feature extraction and retrieval |
| CN108346154A (en) * | 2018-01-30 | 2018-07-31 | 浙江大学 | The method for building up of Lung neoplasm segmenting device based on Mask-RCNN neural networks |
| CN108416780A (en) * | 2018-03-27 | 2018-08-17 | 福州大学 | An object detection and matching method based on twin-region-of-interest pooling model |
| CN108520218A (en) * | 2018-03-29 | 2018-09-11 | 深圳市芯汉感知技术有限公司 | A kind of naval vessel sample collection method based on target tracking algorism |
| CN108846358A (en) * | 2018-06-13 | 2018-11-20 | 浙江工业大学 | Target tracking method for feature fusion based on twin network |
| CN109063574A (en) * | 2018-07-05 | 2018-12-21 | 顺丰科技有限公司 | A kind of prediction technique, system and the equipment of the envelope frame based on deep neural network detection |
| CN109166136A (en) * | 2018-08-27 | 2019-01-08 | 中国科学院自动化研究所 | Target object follower method of the mobile robot based on monocular vision sensor |
| CN109190635A (en) * | 2018-07-25 | 2019-01-11 | 北京飞搜科技有限公司 | Target tracking method, device and electronic equipment based on classification CNN |
| CN109215058A (en) * | 2018-09-17 | 2019-01-15 | 北京云测信息技术有限公司 | A kind of mask method for image recognition face tracking |
| CN109271927A (en) * | 2018-09-14 | 2019-01-25 | 北京航空航天大学 | A kind of collaboration that space base is multi-platform monitoring method |
| CN109543559A (en) * | 2018-10-31 | 2019-03-29 | 东南大学 | Method for tracking target and system based on twin network and movement selection mechanism |
| CN110111363A (en) * | 2019-04-28 | 2019-08-09 | 深兰科技（上海）有限公司 | A kind of tracking and equipment based on target detection |
| CN110322475A (en) * | 2019-05-23 | 2019-10-11 | 北京中科晶上科技股份有限公司 | A kind of sparse detection method of video |
| CN110334635A (en) * | 2019-06-28 | 2019-10-15 | Oppo广东移动通信有限公司 | Subject tracking method, apparatus, electronic device and computer-readable storage medium |
| CN110378938A (en) * | 2019-06-24 | 2019-10-25 | 杭州电子科技大学 | A kind of monotrack method based on residual error Recurrent networks |
| CN110399823A (en) * | 2019-07-18 | 2019-11-01 | Oppo广东移动通信有限公司 | Subject tracking method and apparatus, electronic device, and computer-readable storage medium |
| CN110443824A (en) * | 2018-05-02 | 2019-11-12 | 北京京东尚科信息技术有限公司 | Method and apparatus for generating information |
| CN110472594A (en) * | 2019-08-20 | 2019-11-19 | 腾讯科技（深圳）有限公司 | Method for tracking target, information insertion method and equipment |
| CN110490084A (en) * | 2019-07-24 | 2019-11-22 | 顺丰科技有限公司 | Detection method, device, the network equipment and the storage medium of target object |
| CN110503095A (en) * | 2019-08-27 | 2019-11-26 | 中国人民公安大学 | Alignment quality evaluation method, localization method and the equipment of target detection model |
| CN110634155A (en) * | 2018-06-21 | 2019-12-31 | 北京京东尚科信息技术有限公司 | Target detection method and device based on deep learning |
| CN110634148A (en) * | 2018-06-21 | 2019-12-31 | 北京京东尚科信息技术有限公司 | Method and device for extracting target in continuous frame image |
| CN110647818A (en) * | 2019-08-27 | 2020-01-03 | 北京易华录信息技术股份有限公司 | Identification method and device for shielding target object |
| CN110677585A (en) * | 2019-09-30 | 2020-01-10 | Oppo广东移动通信有限公司 | Target detection frame output method and device, terminal and storage medium |
| CN111462240A (en) * | 2020-04-08 | 2020-07-28 | 北京理工大学 | A target localization method based on multi-monocular vision fusion |
| CN111462174A (en) * | 2020-03-06 | 2020-07-28 | 北京百度网讯科技有限公司 | Multi-target tracking method and device and electronic equipment |
| CN111508006A (en) * | 2020-04-23 | 2020-08-07 | 南开大学 | Moving target synchronous detection, identification and tracking method based on deep learning |
| CN112085762A (en) * | 2019-06-14 | 2020-12-15 | 福建天晴数码有限公司 | Target position prediction method based on curvature radius and storage medium |
| CN112183675A (en) * | 2020-11-10 | 2021-01-05 | 武汉工程大学 | Twin network-based tracking method for low-resolution target |
| WO2021022643A1 (en) * | 2019-08-08 | 2021-02-11 | 初速度（苏州）科技有限公司 | Method and apparatus for detecting and tracking target in videos |
| CN112560651A (en) * | 2020-12-09 | 2021-03-26 | 燕山大学 | Target tracking method and device based on combination of depth network and target segmentation |
| CN112597795A (en) * | 2020-10-28 | 2021-04-02 | 丰颂教育科技(江苏)有限公司 | Visual tracking and positioning method for motion-blurred object in real-time video stream |
| CN112750145A (en) * | 2019-10-30 | 2021-05-04 | 中国电信股份有限公司 | Target detection and tracking method, device and system |
| WO2021208251A1 (en) * | 2020-04-15 | 2021-10-21 | 上海摩象网络科技有限公司 | Face tracking method and face tracking device |
| CN113689462A (en) * | 2020-05-19 | 2021-11-23 | 深圳绿米联创科技有限公司 | Target processing method and device and electronic equipment |
| CN114387296A (en) * | 2022-01-07 | 2022-04-22 | 深圳市金溢科技股份有限公司 | Target track tracking method and device, computer equipment and storage medium |
| CN114612512A (en) * | 2022-03-10 | 2022-06-10 | 豪威科技(上海)有限公司 | KCF-based target tracking algorithm |
| CN116883915A (en) * | 2023-09-06 | 2023-10-13 | 常州星宇车灯股份有限公司 | Target detection method and system based on front and rear frame image association |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104036250A (en) * | 2014-06-16 | 2014-09-10 | 上海大学 | Video pedestrian detecting and tracking method |
| CN105160400A (en) * | 2015-09-08 | 2015-12-16 | 西安交通大学 | L21 norm based method for improving convolutional neural network generalization capability |
| EP3032462A1 (en) * | 2014-12-09 | 2016-06-15 | Ricoh Company, Ltd. | Method and apparatus for tracking object, and non-transitory computer-readable recording medium |
| CN105898107A (en) * | 2016-04-21 | 2016-08-24 | 北京格灵深瞳信息技术有限公司 | Target object snapping method and system |
| CN106022220A (en) * | 2016-05-09 | 2016-10-12 | 西安北升信息科技有限公司 | Method for performing multi-face tracking on participating athletes in sports video |
-
2017
- 2017-01-22 CN CN201710053918.6A patent/CN106875425A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104036250A (en) * | 2014-06-16 | 2014-09-10 | 上海大学 | Video pedestrian detecting and tracking method |
| EP3032462A1 (en) * | 2014-12-09 | 2016-06-15 | Ricoh Company, Ltd. | Method and apparatus for tracking object, and non-transitory computer-readable recording medium |
| CN105160400A (en) * | 2015-09-08 | 2015-12-16 | 西安交通大学 | L21 norm based method for improving convolutional neural network generalization capability |
| CN105898107A (en) * | 2016-04-21 | 2016-08-24 | 北京格灵深瞳信息技术有限公司 | Target object snapping method and system |
| CN106022220A (en) * | 2016-05-09 | 2016-10-12 | 西安北升信息科技有限公司 | Method for performing multi-face tracking on participating athletes in sports video |
Non-Patent Citations (1)
| Title |
|---|
| 毛宁等: "基于分层卷积特征的自适应目标跟踪", 《激光与光电子学进展》 * |
Cited By (68)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107274433A (en) * | 2017-06-21 | 2017-10-20 | 吉林大学 | Method for tracking target, device and storage medium based on deep learning |
| CN107274433B (en) * | 2017-06-21 | 2020-04-03 | 吉林大学 | Target tracking method, device and storage medium based on deep learning |
| CN107330920A (en) * | 2017-06-28 | 2017-11-07 | 华中科技大学 | A kind of monitor video multi-target tracking method based on deep learning |
| CN107330920B (en) * | 2017-06-28 | 2020-01-03 | 华中科技大学 | Monitoring video multi-target tracking method based on deep learning |
| CN107993250A (en) * | 2017-09-12 | 2018-05-04 | 北京飞搜科技有限公司 | A kind of fast multi-target pedestrian tracking and analysis method and its intelligent apparatus |
| CN107730536A (en) * | 2017-09-15 | 2018-02-23 | 北京飞搜科技有限公司 | A kind of high speed correlation filtering object tracking method based on depth characteristic |
| CN107730536B (en) * | 2017-09-15 | 2020-05-12 | 苏州飞搜科技有限公司 | High-speed correlation filtering object tracking method based on depth features |
| CN107730538A (en) * | 2017-10-11 | 2018-02-23 | 恩泊泰（天津）科技有限公司 | A kind of method and device of the multiple target tracking based on image |
| CN107845105A (en) * | 2017-10-24 | 2018-03-27 | 深圳市圆周率软件科技有限责任公司 | A kind of monitoring method, smart machine and storage medium based on the linkage of panorama rifle ball |
| CN107845105B (en) * | 2017-10-24 | 2021-09-10 | 深圳市圆周率软件科技有限责任公司 | Monitoring method based on panoramic gun-ball linkage, intelligent device and storage medium |
| CN107886120A (en) * | 2017-11-03 | 2018-04-06 | 北京清瑞维航技术发展有限公司 | Method and apparatus for target detection tracking |
| CN107992826A (en) * | 2017-12-01 | 2018-05-04 | 广州优亿信息科技有限公司 | A kind of people stream detecting method based on the twin network of depth |
| CN108196680A (en) * | 2018-01-25 | 2018-06-22 | 盛视科技股份有限公司 | Robot vision following method based on human body feature extraction and retrieval |
| CN108196680B (en) * | 2018-01-25 | 2021-10-08 | 盛视科技股份有限公司 | Robot vision following method based on human body feature extraction and retrieval |
| CN108346154B (en) * | 2018-01-30 | 2021-09-07 | 浙江大学 | Establishment method of lung nodule segmentation device based on Mask-RCNN neural network |
| CN108346154A (en) * | 2018-01-30 | 2018-07-31 | 浙江大学 | The method for building up of Lung neoplasm segmenting device based on Mask-RCNN neural networks |
| CN108090918A (en) * | 2018-02-12 | 2018-05-29 | 天津天地伟业信息系统集成有限公司 | A kind of Real-time Human Face Tracking based on the twin network of the full convolution of depth |
| CN108416780A (en) * | 2018-03-27 | 2018-08-17 | 福州大学 | An object detection and matching method based on twin-region-of-interest pooling model |
| CN108416780B (en) * | 2018-03-27 | 2021-08-31 | 福州大学 | An Object Detection and Matching Method Based on Siamese-Region of Interest Pooling Model |
| CN108520218A (en) * | 2018-03-29 | 2018-09-11 | 深圳市芯汉感知技术有限公司 | A kind of naval vessel sample collection method based on target tracking algorism |
| CN110443824A (en) * | 2018-05-02 | 2019-11-12 | 北京京东尚科信息技术有限公司 | Method and apparatus for generating information |
| CN108846358B (en) * | 2018-06-13 | 2021-10-26 | 浙江工业大学 | Target tracking method for feature fusion based on twin network |
| CN108846358A (en) * | 2018-06-13 | 2018-11-20 | 浙江工业大学 | Target tracking method for feature fusion based on twin network |
| CN110634148A (en) * | 2018-06-21 | 2019-12-31 | 北京京东尚科信息技术有限公司 | Method and device for extracting target in continuous frame image |
| CN110634155A (en) * | 2018-06-21 | 2019-12-31 | 北京京东尚科信息技术有限公司 | Target detection method and device based on deep learning |
| CN109063574B (en) * | 2018-07-05 | 2021-04-23 | 顺丰科技有限公司 | Method, system and equipment for predicting envelope frame based on deep neural network detection |
| CN109063574A (en) * | 2018-07-05 | 2018-12-21 | 顺丰科技有限公司 | A kind of prediction technique, system and the equipment of the envelope frame based on deep neural network detection |
| CN109190635A (en) * | 2018-07-25 | 2019-01-11 | 北京飞搜科技有限公司 | Target tracking method, device and electronic equipment based on classification CNN |
| CN109166136B (en) * | 2018-08-27 | 2022-05-03 | 中国科学院自动化研究所 | Target object following method for mobile robot based on monocular vision sensor |
| CN109166136A (en) * | 2018-08-27 | 2019-01-08 | 中国科学院自动化研究所 | Target object follower method of the mobile robot based on monocular vision sensor |
| CN109271927A (en) * | 2018-09-14 | 2019-01-25 | 北京航空航天大学 | A kind of collaboration that space base is multi-platform monitoring method |
| CN109271927B (en) * | 2018-09-14 | 2020-03-27 | 北京航空航天大学 | A Collaborative Monitoring Method for Space-Based Multi-Platforms |
| CN109215058A (en) * | 2018-09-17 | 2019-01-15 | 北京云测信息技术有限公司 | A kind of mask method for image recognition face tracking |
| CN109543559B (en) * | 2018-10-31 | 2021-12-28 | 东南大学 | Target tracking method and system based on twin network and action selection mechanism |
| CN109543559A (en) * | 2018-10-31 | 2019-03-29 | 东南大学 | Method for tracking target and system based on twin network and movement selection mechanism |
| CN110111363A (en) * | 2019-04-28 | 2019-08-09 | 深兰科技（上海）有限公司 | A kind of tracking and equipment based on target detection |
| CN110322475A (en) * | 2019-05-23 | 2019-10-11 | 北京中科晶上科技股份有限公司 | A kind of sparse detection method of video |
| CN112085762A (en) * | 2019-06-14 | 2020-12-15 | 福建天晴数码有限公司 | Target position prediction method based on curvature radius and storage medium |
| CN112085762B (en) * | 2019-06-14 | 2023-07-07 | 福建天晴数码有限公司 | Target position prediction method based on curvature radius and storage medium |
| CN110378938A (en) * | 2019-06-24 | 2019-10-25 | 杭州电子科技大学 | A kind of monotrack method based on residual error Recurrent networks |
| CN110334635A (en) * | 2019-06-28 | 2019-10-15 | Oppo广东移动通信有限公司 | Subject tracking method, apparatus, electronic device and computer-readable storage medium |
| CN110334635B (en) * | 2019-06-28 | 2021-08-31 | Oppo广东移动通信有限公司 | Subject tracking method, apparatus, electronic device, and computer-readable storage medium |
| CN110399823A (en) * | 2019-07-18 | 2019-11-01 | Oppo广东移动通信有限公司 | Subject tracking method and apparatus, electronic device, and computer-readable storage medium |
| CN110490084B (en) * | 2019-07-24 | 2022-07-08 | 顺丰科技有限公司 | Target object detection method and device, network equipment and storage medium |
| CN110490084A (en) * | 2019-07-24 | 2019-11-22 | 顺丰科技有限公司 | Detection method, device, the network equipment and the storage medium of target object |
| WO2021022643A1 (en) * | 2019-08-08 | 2021-02-11 | 初速度（苏州）科技有限公司 | Method and apparatus for detecting and tracking target in videos |
| CN110472594A (en) * | 2019-08-20 | 2019-11-19 | 腾讯科技（深圳）有限公司 | Method for tracking target, information insertion method and equipment |
| CN110472594B (en) * | 2019-08-20 | 2022-12-06 | 腾讯科技（深圳）有限公司 | Target tracking method, information insertion method and equipment |
| CN110647818A (en) * | 2019-08-27 | 2020-01-03 | 北京易华录信息技术股份有限公司 | Identification method and device for shielding target object |
| CN110503095B (en) * | 2019-08-27 | 2022-06-03 | 中国人民公安大学 | Positioning quality evaluation method, positioning method and device of target detection model |
| CN110503095A (en) * | 2019-08-27 | 2019-11-26 | 中国人民公安大学 | Alignment quality evaluation method, localization method and the equipment of target detection model |
| CN110677585A (en) * | 2019-09-30 | 2020-01-10 | Oppo广东移动通信有限公司 | Target detection frame output method and device, terminal and storage medium |
| CN112750145A (en) * | 2019-10-30 | 2021-05-04 | 中国电信股份有限公司 | Target detection and tracking method, device and system |
| CN111462174A (en) * | 2020-03-06 | 2020-07-28 | 北京百度网讯科技有限公司 | Multi-target tracking method and device and electronic equipment |
| CN111462174B (en) * | 2020-03-06 | 2023-10-31 | 北京百度网讯科技有限公司 | Multi-target tracking method and device and electronic equipment |
| CN111462240B (en) * | 2020-04-08 | 2023-05-30 | 北京理工大学 | A target localization method based on multi-monocular vision fusion |
| CN111462240A (en) * | 2020-04-08 | 2020-07-28 | 北京理工大学 | A target localization method based on multi-monocular vision fusion |
| WO2021208251A1 (en) * | 2020-04-15 | 2021-10-21 | 上海摩象网络科技有限公司 | Face tracking method and face tracking device |
| CN111508006A (en) * | 2020-04-23 | 2020-08-07 | 南开大学 | Moving target synchronous detection, identification and tracking method based on deep learning |
| CN113689462A (en) * | 2020-05-19 | 2021-11-23 | 深圳绿米联创科技有限公司 | Target processing method and device and electronic equipment |
| CN112597795A (en) * | 2020-10-28 | 2021-04-02 | 丰颂教育科技(江苏)有限公司 | Visual tracking and positioning method for motion-blurred object in real-time video stream |
| CN112183675A (en) * | 2020-11-10 | 2021-01-05 | 武汉工程大学 | Twin network-based tracking method for low-resolution target |
| CN112183675B (en) * | 2020-11-10 | 2023-09-26 | 武汉工程大学 | A tracking method for low-resolution targets based on Siamese network |
| CN112560651A (en) * | 2020-12-09 | 2021-03-26 | 燕山大学 | Target tracking method and device based on combination of depth network and target segmentation |
| CN114387296A (en) * | 2022-01-07 | 2022-04-22 | 深圳市金溢科技股份有限公司 | Target track tracking method and device, computer equipment and storage medium |
| CN114612512A (en) * | 2022-03-10 | 2022-06-10 | 豪威科技(上海)有限公司 | KCF-based target tracking algorithm |
| CN116883915A (en) * | 2023-09-06 | 2023-10-13 | 常州星宇车灯股份有限公司 | Target detection method and system based on front and rear frame image association |
| CN116883915B (en) * | 2023-09-06 | 2023-11-21 | 常州星宇车灯股份有限公司 | Target detection method and system based on front and rear frame image association |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106875425A (en) | A kind of multi-target tracking system and implementation method based on deep learning | |
| Zheng et al. | Deep learning for event-based vision: A comprehensive survey and benchmarks | |
| CN109387204B (en) | Synchronous positioning and composition method of mobile robot for indoor dynamic environment | |
| Zhou et al. | Efficient road detection and tracking for unmanned aerial vehicle | |
| CN106462976B (en) | Method for tracking shape in scene observed by asynchronous sensor | |
| EP2858008B1 (en) | Target detecting method and system | |
| CN102542289B (en) | Pedestrian volume statistical method based on plurality of Gaussian counting models | |
| CN104200494B (en) | Real-time visual target tracking method based on light streams | |
| CN109934065B (en) | Method and device for gesture recognition | |
| CN110533687B (en) | Multi-target three-dimensional track tracking method and device | |
| CN107958479A (en) | A kind of mobile terminal 3D faces augmented reality implementation method | |
| CN105574894B (en) | A kind of screening technique and system of moving object feature point tracking result | |
| WO2018152214A1 (en) | Event-based feature tracking | |
| CN105913028A (en) | Face tracking method and face tracking device based on face++ platform | |
| CN106780564A (en) | A kind of anti-interference contour tracing method based on Model Prior | |
| CN106157329A (en) | A kind of adaptive target tracking method and device | |
| Min et al. | Coeb-slam: A robust vslam in dynamic environments combined object detection, epipolar geometry constraint, and blur filtering | |
| CN113689459A (en) | Real-time tracking and mapping method based on GMM combined with YOLO in dynamic environment | |
| CN106529441A (en) | Fuzzy boundary fragmentation-based depth motion map human body action recognition method | |
| CN113781523B (en) | A football detection and tracking method and device, electronic equipment, and storage medium | |
| Madasu et al. | Estimation of vehicle speed by motion tracking on image sequences | |
| Pundlik et al. | Time to collision and collision risk estimation from local scale and motion | |
| CN111179281A (en) | Human body image extraction method and human action video extraction method | |
| CN113780181A (en) | Offside judgment method, device and electronic equipment in football match based on drone | |
| Martin et al. | An evaluation of different methods for 3d-driver-body-pose estimation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170620 |